The low cost of human preference incoherence — LessWrong